International Journal of Computer Applications® (IJCA) (0975 – 8887)
National Conference cum Workshop on Bioinformatics and Computational Biology, NCWBCB- 2014
An Improved Chess Machine based on Artificial Neural
Networks
Diwas Sharma
CSE Department
Sikkim Manipal Institute of
Technology
Sikkim, India
ABSTRACT
Numerous published studies revealed that various researchers
have attempted to build a program that learns to play
cognitive games, given little or no earlier knowledge about the
rule of the game. A usual chess playing machine thoroughly
explores the moving possibilities from a chessboard
configuration to choose what the next best move to make. The
brute- force searching technique used by the Deep Blue chess
engine has made vast impact in the ground of artificial
intelligence, but still found to be resource hungry. This paper,
with the concept of Artificial Neural Networks presents a very
simple and efficient approach to develop an intelligent chess
engine which can assist and hint the possible move within the
game using the evolutionary and a adaptive computing
technique on learning from the human grandmasters.
Keywords
Artificial Neural Network, Championship game, En passant,
legal move, Octavius.
1. INTRODUCTION
Chess, known as the game of “perfect information”, since
both players are conscious about the entire situation of the
game at all time, just by looking at the board, has since been
considered as standard of testing intelligence. Artificially
intelligent programs have been developed by such computing
majors as IBM, specifically for the purpose of playing chess,
and after Deep Blue, (Mid 1990’s) [1], defeated the world
chess champion Garry Kasparov, machines have been
consistently beating man in the rapid version of the game. The
kind of victory of a machine over a human grand master has
led to make chess one of the most used games to promote
artificial intelligence and Computer Science. Nevertheless,
chess engines are not able to strategically plan moves or to
explain why they compute such a combination of moves.
This limitation in the power of the machine exists due to the
fact that the decision tree based hardware and software takes
too much time to search the exhaustive option set for the
correct move.
It is found that Neural Networks are rarely used to solve board
games. Major focus is given only to the end-game stages of
the game but few uses of Neural Network in computer chess
game as a whole. The playing strength of the top human chess
players is due to their strategic planning ability, look ahead,
intuition and creativity. The playing strength of the
computers, however, is mainly based on their speed. It would
be more interesting to find out if machine can inherit the way
human player plans the game. There has been little work
published looking at Artificial Neural Network learning where
external(human) knowledge applied to deterministic problems
with large state spaces. This work is interested in investigating
this area by testing whether Artificial Neural Network is
Udit Kr. Chakraborty
CSE Department
Sikkim Manipal Institute of
Technology
Sikkim, India
capable of deriving any useful information in an extremely
sparse environment using a purist approach of introducing
knowledge. In order to achieve this objective the board game
chess was selected as the problem domain.
2. SURVEY
The classical approaches to chess playing by machines have
evolved over the years. Nicolas Lassabe et al [7] mentioned
on their work that, the earliest approach was to evaluate and to
store, for each chessboard configuration the best move to
play. This simplistic and idealistic approach required huge
comparisons and evaluations. Von Neumann and Morgenstern
(1944) proposed minimax as a method to decide which move
to make in chess.[8] A two player game such as chess can be
represented by a tree. Each node is a position on the board,
starting at the root with the beginning position of the game.
The vertices then are the possible moves, leading to the next
nodes (positions). The minimax algorithm is the basic search
tool used in most game programs. Minimax works by building
a depth-first, left-to-right, and search tree with alternating
moves by Max and Min. Only one Artificial Neural Network
based approach for developing a chess engine is reported in
literature by the name of Octavius. Octavius attempts to gain
an understanding of chess through a positional analysis of
master and grandmaster games. Its training is based on the
assumption that any position reached during such games must
be positionally superior to any alternatively available position
during that game. [5]
3. PROBLEM DEFINITION
To generate a function f, which take as input a chess board
position denoted by [13],
Such that,
(1)
Here, X and Y denotes white and black pieces respectively, i
= 1,2,…,16 represents total number pieces for each
side,k=1,2,…,8 denotes files and j=1,2,…,8 denotes ranks,Xi
Bkj/YiBkj represents the current position of X (white) and Y
(black) pieces accordingly.
4. METHODOLOGY
The Figure 1 illustrates the sequence of flow during training
phase where input moves taken from the grandmaster game is
encoded using the board encoding scheme discussed later.
The encoded moves are then fed into the designed neural
network where the system is trained to produce the output.
The system learns the patterns (moves and strategy) that are
generated by the human grandmaster during their
championship games. The system is trained until it produces
the result with minimum error. The Figure 02 shows the flow
8
 International Journal of Computer Applications® (IJCA) (0975 – 8887)
National Conference cum Workshop on Bioinformatics and Computational Biology, NCWBCB- 2014
of designed system during testing phase. Here the trained
system is tested with the moves that are not trained earlier i.e.
user fed unknown moves in unknown environment, where the
system behaves as a move generator. The moves so generated
are validated to check whether they are the valid moves or
not. It also checks the suggested move to meet the rules of the
game.[13]
Fig 1: Training Phase Fig 2: Testing Phase
5. IMPLEMENTATION
Algebraic notation is a method for recording and describing
the moves in a game of chess. Each square of
the chessboard is identified by a unique coordinate pair, a
letter and a number. Each square has a unique identification of
file letter followed by rank number. [14] Considering the fact
that the Artificial Neural Network only understand the
numeric input values those algebraic notations are to be pre-
processed. Firstly, the ANN understandable board is to be
obtained where each square in the chessboard is represented
with the unique ANN understandable numerals, the idea here
is the use of numerical notational technique of the board as
shown in the figure 3,
Fig 3: ANN understandable chessboard[13]
Table 1. Encoding scheme for ANN representation
White Piece Black Piece
Algebraic Numeric Algebraic Numeric
e4 5254 d5 4745
Nf6 7163 Nf3 7866
In the numeric notation all the squares are numbered with a
unique two-digit numbers. In this simple numeric coordinate
system the file and the rank of the chess board is represented
by the numerals 1 to 8 starting from left to right (white side)
and bottom to top respectively. Hence, the individual square
of the chess board can now be denoted by a uniquely
identifiable co-ordinate value where first digit describes
the file and the second one the rank. Hence, the move that
would be written is as shown in the Figure 5.2.The encoded
moves depict the combination of the current position of the
piece as well as the next position to be moved together within
the board. As shown in the Figure 4, the move e4 is
represented as 5254 indicating that a piece is to be moved
from position 52 to next position 54.The work tries to train the
system by means of the patterns generated by the master and
grandmaster during the championship games. [13]
Fig 4: Move e4 in Algebraic and Numerical notation 5254
Fig 5: Proposed Neural Network Structure[13]
The use of adaptive Neural Network techniques allows us to
train the system to construct a winning position, as well as to
develop its pieces which result with the best possible move.
The trained system is then deployed in the testing
environment where the result will be examined based on its
performance. The network structure is of high importance in
such ANN based approaches, and the encoding of the problem
into ANN understandable format is essential. The strategy is
to have 16 inputs and 16 output system where each input
connections signifies a piece of the one player (in order) and
each output connection stands for a piece of the other player.
Each input will be active for only one input, showing a
particular piece movement and have values showing from and
to squares in the chess board. The Network Object created is
then trained using the encoded data that are actually generated
by the human Grandmaster (GM) during Championship
games. The training process requires a set of examples of
proper network behavior - network inputs and the target
outputs. During training the weights and biases of the network
are iteratively adjusted to minimize the error of the network.
The simulation of the designed system allows checking of
performance for the new input moves. The trained network
object is simulated in the new unknown environment where
with the aid of patterns provided to it. This is to test whether it
has learnt to produce an output that is intelligent enough.
6. Results and Discussions
6.1 Test Case I: Scenario (Initial phase of
the game)
The trained system when tested in the unknown environment
found to behave intelligently and was capable enough to
produce the result which drives the chess pieces within the
board on account of the learned patterns. The system when
tested generated the exact moves as the way grandmasters
9
 International Journal of Computer Applications® (IJCA) (0975 – 8887)
National Conference cum Workshop on Bioinformatics and Computational Biology, NCWBCB- 2014

opens the game during the initial stage of the game. The 6.3 Testing of the system in trained
standard patterns followed by the human grandmasters were
clearly imitated by the trained neural network in the early environment:
phase of the game. The designed system when tested with the sequence of trained
move produced the result as shown below in the Table 4. It is
Table 2. Test Data (Scenario 1) found that system behaved very well for the trained sequence
of moves.
Sl. Input Output Remarks
Table 4. Testing result (trained environment)
1 5254 5755 Optimal move
2 2133 7866 Optimal move
3 7163 2836 Optimal move

Sl. Input Output Remarks

1 4244 7866 Good move
Fig 6: Observation Table 2
2 3234 7776 Good move
6.2 Test Case II: Scenario (Intermediate 3 6263 3735 Good move
phase of the game) 4 4445 4746 Good move
After the game progresses, the pieces in the chess board
5 5254 6877 Good move
develop resulting on exponential growth in possibility for
each of the piece to move within the board. During such 6 7152 5870,8839 Castling, positions hinted
scenario the result of the system is found to be very simple 7 5233 6685 Good move
and effective as it realizes and hinted the intermediate move
of a particular piece on the board to be moved next. When an 8 3175 7766 Good move
unknown input move is supplied to the system it is found to 9 7566 5766 Good move
be capable of deciding the position from where the piece is to
10 4142 6665 Good move
be moved. Since the piece to be moved next was clearly
hinted, the tree search technique can help to find out the 11 5465 3865 Good move
suitable destination depending upon the situation of the game 12 7274 6858 Good move
on the board. The Artificial neural network and the tree based
searching technique together in hybrid form is hence found to 13 5141 6521 Good move
be effective that reduce the cost to produce move in terms of 14 1121 4866 Good move
move searching time. Instead of deriving all the possibility
and then searching the entire tree for best possible move, a 15 7485 6663 Good move
tree can formed from the position that is suggested by the 16 4132 6381 Queen thrown
designed system. Fig 8: Observation Table 4
Table 3. Testing result (scenario 2)
6.4 Testing of the system in unknown
Sl. Input Output Remarks environment
The designed system when tested in the new environment
1 6143 3639 Move knight from 36
produced the results that are very much acceptable during
2 6355 4852 Move Queen from 48 initial phase of the game by hinting some of the correct
position of the piece which is to be moved next (eg. 5755).
When the bishop was moved from position 61 to 34, the
system hinted to move pawn from position 67, but is found to
be blocked by knight positioned on 66. Even though it hinted
the invalid move but the system was intelligent enough to
suggest the piece to be move next in the unknown
environment. The system performed even more intelligently
as it hinted to move the knight from position 36 to 44 as
because 44 is found to be a good position supported by one of
the pawn residing in the position 55. But, when a move 6344
is tested it hinted to move a bishop from position 38, which is
Fig 7: Searching move using game tree found to be not a good choice. Instead, it could have captured
that knight by its pawn residing in the position 55.

10
 International Journal of Computer Applications® (IJCA) (0975 – 8887)
National Conference cum Workshop on Bioinformatics and Computational Biology, NCWBCB- 2014

The designed system when tested in the untaught environment Table 6. Test data (unknown environment 2)
produced the following results:
Table 5. Testing result (unknown environment 1) Sl. Input Output Observation Remarks

Sl. Input Output Observation Move pawn Optimal

1 5254 5755
from 57 move
1 5254 5755 Good move
Move knight Optimal
2 6134 7866
2 2133 7866 Good move from 78 move

3 7163 2836 Good move Move knight Optimal

3 4243 2836
from 28 move
Move pawn(knight blocks
4 6134 6705
the move) Move bishop
4 7163 6829 Good move
Move knight (very strong from 68
5 4243 3644
move) Move knight
Move pawn (destination not 5 2133 6671 Good move
6 4142 2755 from 66
defined) Move pawn
6 3175 6705 Good move
7 6344 3845 Move bishop (weak move) from 67
Move pawn
7 7584 7759 Good move
from 77
Move pawn
8 8283 8765 Good move
from 87
Move knight Optimal
9 8473 7460
from 74 move

Fig 9: Observation Table 5
The result of the game shown above in the Table 5 was
readjusted using the concept of game tree technique when
required. The network is then further trained with the input
and the correct output so that the system adapts such scenario.
Again, after training the system with the moves where system
found difficult to produce good result, the system produced
results as shown in the Table 6.
Fig 10: Observation Table 6
7. CONCLUSION
The ANN’s can be trained to learn moves as complex as those
of chess games have been long debated upon. Apart from
Octavian there has been no mention of ANN applications in
chess playing. Even Octavius reports degradation in
performance. Our work brings out the fact that ANN’s can
indeed be used for playing, teaching or helping humans in
playing chess. A true minimax search of a game tree may be
expensive since every leaf node must be visited. Since, for a
uniform tree with exactly W moves at each node, there are
WD nodes at the layer of the tree that is D ply from the root.
Because of the depth of the game tree for chess was
potentially too deep to be successful when using a pure
approach, it was found if Artificial Neural Network and tree
searching technique used in hybrid form can substantially
reduce the move searching cost in terms of time and space. If
sufficient number of training patterns is presented to ANN, it
was found that the machine cleverly hints the suitable position
from where the move can be made. This position can be used
to derive a tree which tends to produce effective moves.
In the simplest case the best algorithm is the one that visits
fewest nodes when determining the true value of a tree. It was
found that the scope of chess as a problem domain was too
broad for minimax approach because of the fact that every
leaf node is to be reached for best move. Therefore, the
problem was narrowed by the way ANN performed. A lot of
time can be saved by the use of this smarter approach as
because there is no requirement to visit all leaves. This
technique can be used to device a chess engine that can play
suggest moves, and to improvise its play, game after game,
improving the understanding gained by the human players
during the play.

8. REFERENCES
[1] Murray Campbell, A. Joseph Hoane Jr., Feng-hsiung
Hsu, “Deep Blue”, IBM,August 1, 2001.
[2] Richard Peter Dazeley, “Investigations into Playing
Chess Endgames using Reinforcement Learning”

11
 International Journal of Computer Applications® (IJCA) (0975 – 8887)
National Conference cum Workshop on Bioinformatics and Computational Biology, NCWBCB- 2014
Bachelor of Computing with Honours, University of
Tasmania (October, 2001)
[3] Jeroen W.T. Carolus, “Alpha-Beta with Sibling
Prediction Pruning in Chess” Netherlands 2006
[4] C.E. Shannon, “Programming a Computer for Playing
Chess”, Philosophical Magazine, Ser. 7, Vol. 41, Bell
Telephone Laboratories, March 1950
[5] AlexandruGodescu, “An information theoretic analysis
of decision in computer chess” ETH Zurich December
13, 2011
[6] LaurenceFausett, “Fundamentals of Neural Networks
Architectures, Algorithms, and Applications”, Dorling
Kindersley Pvt. Ltd., South Asia, 2008
[7] Nicolas Lassabe, Stephane Sanchez HerveLuga and Yves
Duthen,“Genetically Programmed Strategies for Chess
Endgames”. GECCO’06, July 8-12, 2006
[8] T.A. Marsland and Y. Björnsson, “From Minimax to
Manhattan,”, University of Alberta Canada
[9] Olivier Mehani, “Complex Systems Seminar, Can
complex systems help design better chess-playing
programs”, February, 1 2006
[10] Robert Epstein, Gary Roberts, Grace Beber, “Parsing the
turing test philosophical and Methodological Issues in
the Quest for the Thinking Computer”. Springer, 03-Dec-
2008
[11] Robert Levinson and Ryan Weber, “Chess
Neighbourhoods, Function Combination and
Reinforcement Learning”, University of California Santa
Cruz, Springer, October 2000
[12] JoonatanManttari and Jonas Larsson, “Applications of
Artificial Neural Networks in Games; An Overview”.
School of Innovation, Design and Engineering
[13] Diwas Sharma, Udit Kr. Chakraborty, Artificial Neural
Network based adaptive chess playing machine.
International Journal of Advance Research in Computer
Science, Vol 4 no. 4, March-April 2013.
[14] http://en.wikipedia.org/wiki/ICCF_numeric_notation
dated: 12/10/2012
[15] Adam Berent “ChessBin” Version: 0.0.8.0 (snapsorts)
12